Re: [PATCH perf/core 03/22] uprobes: Move ref_ctr_offset update out of uprobe_write_opcode

[Oleg Nesterov <oleg@redhat.com>]

On 04/21, Jiri Olsa wrote:
>
> +static int set_swbp_refctr(struct uprobe *uprobe, struct vm_area_struct *vma, unsigned long vaddr)
> +{
> +	struct mm_struct *mm = vma->vm_mm;
> +	int err;
> +
> +	/* We are going to replace instruction, update ref_ctr. */
> +	if (uprobe->ref_ctr_offset) {
> +		err = update_ref_ctr(uprobe, mm, 1);
> +		if (err)
> +			return err;
> +	}
> +
> +	err = set_swbp(&uprobe->arch, vma, vaddr);
> +
> +	/* Revert back reference counter if instruction update failed. */
> +	if (err && uprobe->ref_ctr_offset)
> +		update_ref_ctr(uprobe, mm, -1);
> +	return err;
>  }
...
> +static int set_orig_refctr(struct uprobe *uprobe, struct vm_area_struct *vma, unsigned long vaddr)
> +{
> +	int err = set_orig_insn(&uprobe->arch, vma, vaddr);
> +
> +	/* Revert back reference counter even if instruction update failed. */
> +	if (uprobe->ref_ctr_offset)
> +		update_ref_ctr(uprobe, vma->vm_mm, -1);
> +	return err;
>  }

This doesn't look right even in the simplest case...

To simplify, suppose that uprobe_register() needs to change a single mm/vma
and set_swbp() fails. In this case uprobe_register() calls uprobe_unregister()
which will find the same vma and call set_orig_refctr(). set_orig_insn() will
do nothing. But update_ref_ctr(uprobe, vma->vm_mm, -1) is wrong/unbalanced.

The current code updates ref_ctr after the verify_opcode() check, so it doesn't
have this problem.

-------------------------------------------------------------------------------
OTOH, I think that the current logic is not really correct too,

	/* Revert back reference counter if instruction update failed. */
	if (ret < 0 && is_register && ref_ctr_updated)
		update_ref_ctr(uprobe, mm, -1);

I think that "Revert back reference counter" logic should not depend on
is_register. Otherwise we can have the unbalanced update_ref_ctr(-1) if
uprobe_unregister() fails, then another uprobe_register() comes at the
same address, and after that uprobe_unregister() succeeds.

Oleg.